Page 1
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
Store
Process
Analyze
Collaborate
Archive
Cloud
The HPC Storage Leader Invent
Discover
Compete
1
Page 2
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
DDN | Who We Are
► Main Office: Sunnyvale, California, USA
► Go To Market: Partner & Reseller Assisted, Direct
► DDN: World’s Largest Private Storage Company
► Only Storage Company with Long-Term on Big Data Focus
We Design, Deploy and Optimize Storage Systems Which
Solve HPC, Big Data and Cloud Business Challenges at Scale
World-Renowned & Award-Winning
2
Page 3
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
3 An Elite Collection Of HPC’s Finest... Some of our 1000+ Customers
Page 4
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
DDN | 15 Years of HPC Innovation 4
DDN Leads On The List of Lists:
80% of the Top 10
67% of the Top 100
32% of the Top 500
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
DDN FOUNDED LARGEST PRIVATE
STORAGE CO. (IDC)
500+
EMPLOYES
1st Real-Time Appliance
for High-Scale Big Data 1st in Data Center
Density
1st CUSTOMER
NASA
SFA Storage Fusion Architecture
2013
1st in Bandwidth + IOPS
1st In-Storage Processing™
SW-Only, Portable Architecture DDN’s 1st Parallel File
System Offering ft. Lustre 1st Hyperscale Object
Storage
Web-Scale Computing
and HPC Collaboration
SFX Flash Tiering
Revolutionizing HPC
1st Application-Aware
Hybrid Caching EXAScaler™
2014
10GB/s NCSA
100GB/s CEA, LLNL
1TB/s ORNL
5 BP / Rack
Page 5
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
5 Our Unwavering Commitment to HPC
Investments in Exascale Real Engineering Is Needed To Scale 1000x
Fast Forward
Page 6
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
Exascale I/O Challenges - Cost 6
LANL Trinity Hybrid Scratch Cost Analysis
Hybrid approach is necessary to meet bandwidth & capacity requirements
Page 7
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
Exascale I/O Challenges – Power Consumption 7
7
0
10000
20000
30000
40000
50000
60000
70000
0.76 2.96
# o
f H
DD
s
Burst Throughput (TB/sec)
NERSC-8 Cost Comparison
Hybrid
HDDs
Power – 470KW
26 SFA Controllers
Power – 768KW
26 SFA Controllers + BB
Power – 1792KW
99 SFA Controllers
Page 8
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
Burst Buffer
Absorbs the
Peak Load
Filesystem
Handles the
Remaining Load
Analysis: Argonne’s LCF production storage system (circa 2010) • 99% of the time, storage BW utilization < 33% of max
• 70% of the time, storage BW utilization < 5% of max
Archival
Storage Tier
Persistent
Storage Tier
Burst Buffer
Tier
IME SC’13 Demo
Cluster
25 MB/s
4 GB/s
50 GB/s
1) Separation of bandwidth and capacity is required
2) Utilization efficiency must be improved
8 Exascale I/O Challenges – Efficiency
Page 9
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
Why is today’s I/O efficiency so poor?
► Serialization at various points in the I/O
path
• Stripe and block alignment (PFS and RAID) o Read-modify-writes to underlying storage
• Lock contention o Exacerbated by poor I/O structuring in applications
File Server 1 File Server 2 File Server 3 File Server 4
Storage
Compute
Node1
Compute
Node2
Lock Contention
Worsens with 1000s of nodes
2015 2018
Performance(TF)
20000 1000000
Concurrency 5000000 1000000000
0
200000000
400000000
600000000
800000000
1E+09
1.2E+09
1000
10000
100000
1000000
10000000
100000000
Source: http://storageconference.org/2011/Presentations/SNAPI/1.Grider.pdf
9
Page 10
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
Why is today’s I/O efficiency so poor?
► Poor horizontal scaling characteristics
in the PFS – weakest link
• PFS are only as fast as the slowest I/O
component
• Oversubscribed or crippled I/O
components affect the entire system
performance
• As I/O sections get larger and # of
components increases the problem
worsens (congestion)
• This weakest link can be all the way down
to disks (RAID rebuilds)
File Server 1 File Server 2 File Server 3 File Server 4
Storage
A single overloaded
server can slow
down the entire
system
10
Page 11
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
PFS Efficiency as a Function of I/O Size 11
0
10
20
30
40
50
60
70
80
90
100
0
10
20
30
40
50
60
70
80
90
100
4096 409600
Perc
en
tag
e o
f S
trip
e S
ize
Pe
rfo
rma
nc
e E
ffic
ien
cy (
Pe
rce
nt)
I/O Size (Log-Scale)
Performance & Efficiency of Non-Mergeable Writes as a
Function of I/O Size
Performance
I/O Size (bytes)
0
200
400
600
800
1000
1200
1400
1600
1800
2000
1 8 64 512
Th
rou
gh
pu
t (M
B/s
)
IO Request SIze (KB)
Parallel Filesystem on IME Demo Cluster SSDs (50GB/s available)
Avg…
Aligned, full-stripe-width IO required
for maximum PFS I/O performance
Faster media (SSDs) may not address
the underlying PFS performance limitations
Page 12
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
What is Infinite Memory Engine (IME™)?
High performance I/O system based on
parallel log structuring
► Massive concurrency regardless of
application I/O pattern
► Dynamically load balancing helps steer clear
of oversubscribed and handicapped
components
► Innovative lookup mechanism enables
immediate availability of data
► Distributed fault tolerance
12
12
Page 13
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
13 The Infinite Memory Advantage
Designed for Scalability
Patented DDN Algorithms
Scale-Out Data Protection
Distributed Erasure Coding
Integrated With File Systems
Designed to Accelerate Lustre*,
GPFS
No Code Modification Needed
Fully POSIX & HPC Compatible
No Application Modifications
Non-Deterministic System
Write Anywhere, No Layout Needed
Writes: Fast. Reads: They’re Fast Too.
No other system offers both at scale.
Page 14
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
14 SC‘13 Demo Comparative Testing: Shared
Writes 14 IME Clients one per compute node; 98 node-local MLC SSDs
DISK LEVEL TESTING DDN GRIDScaler™
(per SSD)
IME
(per SSD)
62.5 Concurrent Write Requests 438 MB/s 500 MB/s
125,000 Concurrent Write Requests 170 KB/s 500 MB/s
CLUSTER LEVEL TESTING DDN GRIDScaler™ IME( overall)
6,225 Concurrent Write Requests
(8 MB)
49 GB/s 49 GB/s
12,250,000 Concurrent, Interleaved
Write Requests (4 KB)
17 MB/s 49 GB/s
SSDs behind a PFS don’t help
IME is at line rate to scale with SSD rates
Linear Cluster Scaling
Avg. 2018 Top500
Cluster Concurrency
57,772,000 Cores (est)
Page 15
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
IME Checkpoint / Migration Workload Demo
Achieves >90% of Available Storage Bandwidth
• Checkpoint I/O directed at
IME (emulated with IOR)
• File #1 (49 – 50 GB/s)
• File #2 (49 – 50 GB/s)
• File #3 (49 – 50 GB/s)
• Migration of File #3 from IME
to PFS (4 -5 GB/s)
15
Page 16
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
ISC’14 IME Demo Server 16
► Off the shelf 2U Server Chassis
► Dual Socket Ivy Bridge with 128
GB RAM
► Up to 24 SSDs per IME Server
► 2 FDR IB Ports
► Expected Burst Bandwidth per
IME Server: ~10GB/s
Page 17
© 2014 DataDirect Networks, Inc. * Other names and brands may be claimed as the property of others.
Any statements or representations around future events are subject to change. ddn.com
ISC’14 Demo System in DDN Booth 17
► 16U (servers)
► Total Peak BW: ~80GB/s